Re: [PATCH] mm: mglru: clear the reference counter for rejected folios
From: Baolin Wang
Date: Mon Sep 07 2026 - 01:57:54 EST
On 9/7/26 12:28 PM, Barry Song wrote:
On Mon, Sep 7, 2026 at 11:25 AM Baolin Wang
<baolin.wang@xxxxxxxxxxxxxxxxx> wrote:
As per the comment on LRU_REFS_FLAGS, when accessed folios are promoted to
a new generation, LRU_REFS_FLAGS should be cleared so that the reference
counter can start over.
For folios rejected by shrink_folio_list(), we clear LRU_REFS_FLAGS and
set the PG_active flag if the rejected folio is planned to be put back to
the oldest generation. That's fine.
I find the description a bit confusing. Yes, according to
`lru_gen_folio_seq()`, the folio would be put back into the oldest
generation. That's why we set the active bit to prevent this from
happening. So there is no case where the folio is intentionally put
back into the oldest generation.
How about?
"
For folios rejected by shrink_folio_list(), we clear LRU_REFS_FLAGS and
set the PG_active flag when lru_gen_folio_seq() would place them in the oldest generation.
"
But for those that are not put back to the oldest generation (which can
be treated as a promotion), we do not clear LRU_REFS_FLAGS, which can
violate the promotion mechanism. This means the rejected folio enters the
new generation with stale, inflated tier bits, which can inflate reference
counts and distort eviction statistics for these rejected folios.
Maybe simply say that folios for which lru_gen_folio_seq()
returns something other than min_seq?
OK.
Fix this by clearing LRU_REFS_FLAGS for rejected folios, and also do some
measurement. On my 32-core Arm machine, with the memcg limit set to 3G,
running 'make -j32' to build the kernel showed a small improvement in sys
time when using either a zram or NVMe swap device (averaged over 2 runs with
no significant variance).
zram swap:
w/o patch w/ patch
sys time: 1666.5s 1589.5s
NVMe swap:
w/o patch w/patch
sys time: 760s 741.5s
Signed-off-by: Baolin Wang <baolin.wang@xxxxxxxxxxxxxxxxx>
---
mm/vmscan.c | 7 ++++---
1 file changed, 4 insertions(+), 3 deletions(-)
diff --git a/mm/vmscan.c b/mm/vmscan.c
index 40d3f1b48a74..42c0a09938ab 100644
--- a/mm/vmscan.c
+++ b/mm/vmscan.c
@@ -5021,10 +5021,11 @@ static int evict_folios(unsigned long nr_to_scan, struct lruvec *lruvec,
}
/* don't add rejected folios to the oldest generation */
- if (lru_gen_folio_seq(lruvec, folio, false) == min_seq[type]) {
- folio_set_lru_refs(folio, 0);
+ if (lru_gen_folio_seq(lruvec, folio, false) == min_seq[type])
folio_set_active(folio);
- }
+
+ /* See the comments on LRU_REFS_FLAGS */
+ folio_set_lru_refs(folio, 0);
}
The code change makes sense to me if we keep aging and always reclaim
only the oldest generation. However, if we go further in
`scan_folios()`—for example, continuing to scan after we have fewer
than two generations left—we might not be in that case anymore.
I may not fully understand your concern. If there is only one generation (fewer than two generations), we should aging instead of reclaim.
But my point is that regardless, if promotion happens, we should clear the reference counter. Otherwise, as I mentioned, it will distort eviction statistics.
So far, the code looks correct to me.
BTW, Baolin, I wonder if you could find some time to test this
change on your arm64 system:
https://lore.kernel.org/linux-mm/20260901220430.79810-1-baohua@xxxxxxxxxx/
It might have addressed one of your previous, similar concerns here.
On my x86 machine, I find that the change further reduces the
sys time from 3 minutes 40 seconds to 3 minutes 30 seconds.
Meanwhile, I will also run your patch on my x86 machine.
Sigh. I missed this thread. Yes, I also raised a similar issue before [1] and this is still my next step. I'll find some time to check this thread. Thanks for reminding me.
[1] https://lore.kernel.org/all/eb395442-0aad-428a-a5ac-9072d2d89060@xxxxxxxxxxxxxxxxx/